Re: [patch 2/5] x86/kvm: Handle task_work on VMENTER/EXIT

[Peter Zijlstra]

On Thu, Aug 01, 2019 at 02:35:50PM -0700, Sean Christopherson wrote:
> On Thu, Aug 01, 2019 at 08:34:53PM +0200, Thomas Gleixner wrote:
> > On Thu, 1 Aug 2019, Oleg Nesterov wrote:
> > > On 08/01, Thomas Gleixner wrote:
> > > >
> > > > @@ -8172,6 +8174,10 @@ static int vcpu_run(struct kvm_vcpu *vcp
> > > >  			++vcpu->stat.signal_exits;
> > > >  			break;
> > > >  		}
> > > > +
> > > > +		if (notify_resume_pending())
> > > > +			tracehook_handle_notify_resume();
> > > 
> > > shouldn't you drop kvm->srcu before tracehook_handle_notify_resume() ?
> > > 
> > > I don't understand this code at all, but vcpu_run() does this even before
> > > cond_resched().
> > 
> > Yeah, I noticed that it's dropped around cond_resched().
> > 
> > My understanding is that for voluntary giving up the CPU via cond_resched()
> > it needs to be dropped.
> > 
> > For involuntary preemption (CONFIG_PREEMPT=y) it's not required as the
> > whole code section after preempt_enable() is fully preemptible.
> > 
> > Now the 1Mio$ question is whether any of the notify functions invokes
> > cond_resched() and whether that really matters. Paolo?
> 
> cond_resched() is called via tracehook_notify_resume()->task_work_run(),
> and "kernel code can only call cond_resched() in places where it ...
> cannot hold references to any RCU-protected data structures" according to
> https://lwn.net/Articles/603252/.

This is SRCU, you can reschedule while holding that just fine. It will
just delay some kvm operations, like the memslot stuff. I don't think it
is terrible to keep it, but people more versed in KVM might know of a
good reason to drop it anyway.